Hermes may work better with one well-designed agent
Hermes was run on a personal computer with an , using as the main model and as a backup. The divided work among a lead, a fact checker, a writer, and a visual specialist. This separation made it easier to keep checking apart from writing, locate the stage causing a problem, and choose a different AI model for each job.
However, passing work between agents increased waiting time and consumed more of the available . With 24 GB of graphics-card memory, one agent guided by a strong SOUL.md and clear instructions sometimes matched or beat the larger setup by switching roles as needed. Coordination among the agents also failed more often than expected.
It remains unclear whether the extra structure consistently improves output quality or mainly adds complexity.
Key points
- Separate lead, checking, writing, and visual roles make responsibility clearer.
- A different AI model can be assigned to each kind of work.
- Passing tasks between agents uses more time and .
- With 24 GB of graphics-card memory, one well-designed agent sometimes performed as well as or better than several agents.
- Coordination failures can reduce the value of a .